Performance evaluation of neural networks in federated learning for classification of alzheimer’s disease stages

Alzheimer’s disease is a progressive neurodegenerative disorder where early detection is critical for improving treatment outcomes. This study investigates the performance of ResNet 50 and DenseNet 169 architectures integrated with Federated Learning (FL) techniques to classify the stages of Alzheimer’s disease into three main groups: AD (Alzheimer’s Disease), MCI (Mild Cognitive Impairment) and CN (Cognitively Normal). We evaluate three aggregation strategies (FedAvg, FedAvgM and FedProx), across varying client configurations (3-7), while introducing a hybrid FL-Split Learning (SL) approach to optimize computational efficiency and privacy preservation. As a results of the ResNet-FedProx combination achieved superior performance (91.2% accuracy) in 7-client federated scenarios. In contrast, FedAvgM showed unexpected limitations, with accuracy decreasing by 6.7 percentage points as the number of customers increased. Future work should explore hybrid neural networks that combine ResNet and DenseNet architectures and adapt them to other aggregation models available in the literature to optimize performance.